A Practical RBF Framework for Database Load Balancing Prediction

摘要

In this paper, aiming at the load situation of database cluster, we analyze the commonly used load balancing algorithms in the past, but these traditional algorithms fail to meet the fine-grained load prediction of long-term database operation. In order to solve the load balance of the database level in the cloud on the database, we put forward the RBF framework to solve the problem of real-time prediction of database load. The RBF framework we put forward can achieve real-time prediction, and make the prediction of server and even database load. In this process, the delay monitoring information is used to update the model parameters. The main process includes: 1) the index data in the process of collection system and database running forms the input flow. 2) according to the threshold method, the input flow is respectively entered into the positive and negative example pools. When the positive or negative example pools are full, the input flow is added into the backup pool according to the corresponding rules. Discard the data with certain rules. 3) embed the data in the positive case pool, the data in the negative case pool and the data in the backup pool into neurons to form input. 4) using RBF network to forecast the load data in real time. 5) the prediction of incorrect data is formed into delay supervision information, the operation of back to the pool is carried out, and the probability of being selected is increased. 6) calculate the value of fitness function and evaluate whether there is turbulence in the prediction process. Data level load forecasting is fine-grained. In this paper, we simplify the indicators that need to be counted. Compared with other load balancing algorithms, RBF framework is a framework that can carry out long-term load balancing scheduling and forecasting. It is mainly aimed at the load forecasting of database application, and provides a feasible load forecasting for the cloud on the national grid database of China Ideas, and in some database clusters have achieved good results, reached the advanced level.